I’m OK with the upgrade to 4.7 but I agree the harness should be treated more like infrastructure.

My experience was not recent either, everything is so bleeding new that I doubt I could reproduce it

There’s a certain elasticity of demand. The faster tech people can produce tech artefacts that are wanted by non-tech people, the more of those tech artefacts will be wanted.

I think the bottle neck will always be the tech people, even when they have AI tools that make them go super fast.

Oh and I think the reason it’s so obstinate in your case may be that they actually have a copilot for Excel product. I’m not 100% sure on that though.

GitHub copilot (or any dev harness) can make you a Jupyter notebook that can handle xlsx files- side bonus is that you see AI powered data tools :)

@greggmoore.bsky.social This is the immediate challenge after the “ah hah!” moment of building something with AI that seems unbelievably good. That’s where systems and software thinking starts to be needed.

Luckily you can learn these from AI!

I focus on Claude code for interactivity, getting through one topic with like five or six different messages back-and-forth. I would be frustrated that copilot would treat each response as its own hit to my quota.

I use the first tier of both instead of upgrading either. I found that copilot will consume the same for a big query as a small one- I take advantage of that. The way I work, I have agile style stories that are broken down to be arbitrary long. I’ll make a big one and feed it to copilot.

I’m glad I didn’t over invest in protecting this because it’s been over a year now and I’ve literally never needed my old windows OS. if I ever need space, this will be the first thing I clean up :)

I’d love to have been able to convert my windows partition converted into a vm. I ended up setting up dual boot as insurance against needing to get back into my old os.

That said, I haven’t needed it ever… I may wipe the partition

VS code for txt editing, notepad++ for coding!

I’ve used plan execute separation for a long time. To be honest I haven’t found much value in plan mode at all- instead I have a planning agent that understands my workflow organization.

It’s really hard to imagine a good metric that would directly measure what AI coding is giving us. You’d want to leave the technical realm of counting lines of code and such, but that leads you to a very subjective realm of trying to quantify the utility of the code being generated.

AI with Jupyter notebooks feels like a cheat code.

I had a fun experience where even after linking it its own documentation, it wasn’t able to make it work.

GPT and Claude models are both offered through GitHub Copilot. Different tasks favour different models however when I was experimenting I ended up getting better results with Claude.

I suppose the definition of “there” depends on your perspective. I can now reliably build anything I can describe without having to actually write the language I choose.

Have you tried it? In order to be able to understand what it can do, I spent some time using AI development tools to build something in a language I don’t know.

I did spend time following tech debt dead ends, but the constraints that taught me permit useful development now.

I actually don’t use plan mode, but I have it write epics and then stories for those epics. Then in another instance, it reads the story, forms a plan, and does the work.

If you want it pleasing and intuitive, then you want more than a proof of concept right? My problem with Swift was that I couldn’t use it for anything outside of the walled garden.

I ended up using HTML5 with a capacitor based translation- one code base, all platforms.

I’m pretty sure I fail those more often than a language model would these days :(

One possibility that would be cool is that the ability to usefully parse brain signals from different brains might emerge from such a system the way multilingualism emerged from language models.

It’s not the last good one but it’s not listed here and I enjoyed the heck out of it: Discovery

So we train language models using sample language data. I know it’s structurally different but I wonder if there’s any way we could link up brain scan sample data to this language sample data. Make this create something that could generate text responses to your thoughts?

I would spend significant time on a platform that made it hard to automate communication. Something where I’m more likely to be talking to a real person

So I have the AI include evidence that the result of the command “npx jest --coverage” shows full coverage with every PR. Also the project instructions are clear that they’re working with TDD so must write a failing test before writing any code.

I have a quality gate for test coverage. With the language I’m working in, it’s easy to see whether every line is covered by a test or not. Is that what you meant by what is being tested?

Feeds or something like them should be baked in. I shouldn’t have to find somebody willing to host it on their own dime.

We’re at the noisy side of the hype cycle. I predict that eventually there will be plenty of people who can’t code that think of using AI to build software as just a basic capability. Just like I see dividing large sums with lots of decimals to be a trivial task, but my grandparents respected it

People want this to not be for real, it makes them uncomfortable on a philosophical level.